Good Day Sunshine: Stock Returns and the Weather

Page created by Arthur Carter
 
CONTINUE READING
THE JOURNAL OF FINANCE  VOL. LVIII, NO. 3  JUNE 2003

 Good Day Sunshine: Stock Returns and the Weather

                      DAVID HIRSHLEIFER and TYLER SHUMWAY n

                                            ABSTRACT
      Psychological evidence and casual intuition predict that sunny weather is as-
      sociated with upbeat mood. This paper examines the relationship between
      morning sunshine in the city of a country’s leading stock exchange and daily
      market index returns across 26 countries from 1982 to 1997. Sunshine is
      strongly signi¢cantly correlated with stock returns. After controlling for sun-
      shine, rain and snow are unrelated to returns. Substantial use of weather-
      based strategies was optimal for a trader with very low transactions costs.
      However, because these strategies involve frequent trades, fairly modest costs
      eliminate the gains.These ¢ndings are di⁄cult to reconcile with fully rational
      price setting.

SUNSHINE AFFECTS MOOD, as evidenced by song and verse, daily experience, and
formal psychological studies. But does sunlight a¡ect the stock market?
   The traditional e⁄cient markets view says no, with minor quali¢cations. If sun-
light a¡ects the weather, it can a¡ect agricultural and perhaps other weather-
related ¢rms. But in modern economies in which agriculture plays a modest role,
it seems unlikely that whether it is cloudy outside the stock exchange today
should a¡ect the rational price of the nation’s stock market index. (Even in coun-
tries where agriculture plays a large role, it is not clear that one day of sunshine
versus cloud cover at the stock exchange should be very informative about har-
vest yield.)
   An alternative view is that sunlight a¡ects mood, and that people tend to eval-
uate future prospects more optimistically when they are in a good mood than
when they are in a bad mood. A literature in psychology has found that mood
a¡ects judgment and behavior. The psychological literature on sunlight, mood,
and misattribution of mood is discussed in the next section. An important strand
of this literature has provided evidence that mood contains valuable information

   n
     Hirshleifer is from the Fisher College of Business, Ohio State University and Shumway is
from the University of Michigan Business School. We thank Aldo Rosas for his excellent re-
search assistance. We also thank the editor, Rick Green; an anonymous referee; Andrew Ang;
Michael Cooper; Joshua Coval; Bud Gibson; Marc Hulbert; Alice Isen; Daniel Kahneman;
Seongyeon Lim; Barbara Mellers; Marina Murphy; Norbert Schwarz; Lu Zheng; Jason Zweig;
and seminar participants at Cornell University, the Decision Sciences Consortium, and the
Finance Department brown bag seminar at the University of Michigan for helpful comments.
Any remaining errors are our own.

                                                 1009
1010                             The Journal of Finance

about the environment. The inferences drawn from mood can go astray however;
people sometimes attribute their mood to the wrong feature of the environment.
For example, someone who is in a good mood because of sunshine may uncon-
sciously attribute this feeling to generally favorable life prospects. If such misat-
tribution extends to investments, then stock prices will £uctuate in response to
the mood of investors. Furthermore, people in good moods ¢nd positive material
more salient and psychologically available than negative material. This evidence
suggests that on dim, dull, dreary, depressing days stocks will decline, whereas
cheery bright days will boost stocks.
   For three reasons, examining the e¡ects of sunlight on the stock market
provides an attractive means of testing whether psychological biases can a¡ect
stock returns. First, any such relationship is not subject to the criticism of data
snooping. Exploration of whether this pattern exists was speci¢cally stimulated
by the psychological hypothesisFthe hypothesis was not selected to match a
known pattern. Second, such a pattern, if it exists, has a psychological explana-
tion but no plausible rational explanation. This contrasts with many well-known
patterns of stock returns for which psychological and rational explanations are
currently competing.Third, sunshine is an easily measured exogenous in£uence.1
Some previous work emphasizes the dynamics of the social transmission of pop-
ular sentiments and theories (see, e.g., the overview of Shiller (2000a)). Some
empirical headway along these lines has been made through survey methods
(see, e.g., Shiller (1990, 2000b) and Hong, Kubik, and Stein (2003)). Here, we side-
step the complexities of the process of social in£uence by focusing on an exogen-
ous external in£uence.
   The ¢rst test of the hypothesis that sunshine a¡ects returns involves examin-
ing individually for each city the relationship between daily cloudiness and daily
nominal return on the nation’s stock index using univariate regression at 26
stock exchanges internationally from 1982 to 1997. To ensure that the e¡ects we
identify do not derive from well-known seasonal stock return e¡ects, we examine
the deviation between the day’s cloudiness and the ordinary expected degree of
cloudiness for that day of the year.We examine the relationship of cloudiness both
to continuous returns, and (in logit regressions) to the probability that the
return will be positive. Depending on the speci¢cation, in either 18 or 25 of the
26 cities, the relationship between returns and cloudiness is negative (in most
cases not signi¢cantly). Based on a simple nonparametric joint test, it is unlikely
that these results would arise by sheer chance.
   These city-by-city results strongly suggest that there is a genuine relationship
between stock returns and cloudiness.To examine this issue in a more structured
way, we provide several joint (cross-city) parametric tests using the entire data
set. In pooled regressions where the intercept and slope are constrained to be the
same across cities, we ¢nd a highly statistically signi¢cant relationship between

  1
   As Roll (1992) put it, ‘‘Weather is a genuinely exogenous economic factor. y It was a fa-
vourite example of an exogenous identifying variable in the early econometrics literaturey
Because weather is both exogenous and unambiguously observable y weather data should be
useful in assessing the information processing ability of ¢nancial markets.’’
Stock Returns and theWeather                              1011

cloud cover and returns (t 5  4.49), or (w2 5 43.6), for the logit. However, these
tests assume independent errors, which is implausible.
   To address this issue, we estimate a city-speci¢c ¢xed e¡ects model with panel
corrected standard errors (PCSE). Our PCSE speci¢cation allows errors to be
contemporaneously correlated, heteroskedastic across cities and autocorrelated
within each city’s time series. Once again there is a highly signi¢cant negative
relationship between cloud cover and returns (asymptotic z-statistic 5  3.96).
To examine the relationship of weather to the probability of a positive return,
we estimate a ¢xed e¡ects linear probability model with PCSE using an indicator
variable that is one when returns are positive.This again yields a strongly signif-
icant relationship (t 5  6.07). In all cases, adjusting for contemporaneous corre-
lation and heteroskedasticity across panels and autocorrelation within panels
has little e¡ect on the inference.
   The magnitude of the sunshine e¡ect is substantial. For example, for New York
City, the annualized nominal market return on perfectly sunny days is approximately
24.8 percent per year versus 8.7 percent per year on perfectly cloudy days. However,
from an investor’s perspective, the value of these return di¡erentials depends on
whether it is possible to diversify the risk of a sunshine-based trading strategy. We
¢nd that for very low levels of transaction costs, trading strategies based on the
weather generate statistically signi¢cant but economically modest improvements
in portfolio Sharpe ratios. However, some countries’ stock indices cannot be traded
cheaply, and weather strategies involve frequent trading, so for most investors, trad-
ing on the sunshine e¡ect does not appear to be pro¢table net of costs.2
   We also explore by means of multiple regression whether the e¡ect derives
from sunshine versus cloudiness, or from other associated weather conditions
such as rain or snow. We ¢nd that sunshine remains signi¢cant, and that after
controlling for sunshine, rain and snow are essentially unrelated to returns.
   The remainder of the paper is structured as follows. Section I discusses misat-
tribution of mood, sunshine, and stock returns. Section II describes the data we
use for our analyses, while Section III reports our evidence in detail. Section IV
concludes.

          I. Misattribution of Mood, Sunshine, and Stock Returns
A. Mood, Judgment, and Decisions
   A literature in psychology considers how emotions and moods in£uence human
decision making. Individuals who are in good moods make more optimistic
choices. A highly robust e¡ect is that individuals in good moods have more posi-
tive evaluations of many sorts, such as life satisfaction, past events, people,
and consumer products (see, e.g., Wright and Bower (1992), and the survey of
Bagozzi, Gopinath, and Nyer (1999)). There is a mood congruency e¡ect, wherein
  2
    Our ¢ndings do not rule out the possibility that weekly or more complex weather-based
strategies can retain high returns while further economizing on transaction costs; no doubt
practitioners will explore these possibilities.
1012                                The Journal of Finance

people who are in bad moods (good moods) tend to ¢nd negative (positive) mate-
rial more available or salient (see, e.g., Isen et al. (1978) and Forgas and Bower
(1987)). Mood most strongly a¡ects relatively abstract judgments about which
people lack concrete information (Clore, Schwarz, and Conway (1994) and Forgas
(1995)).
   Several studies have found that individuals who are in good moods engage in
more use of simplifying heuristics to aid decisions (see the reviews of Bless,
Schwarz, and Kemmelmeier (1996) and Isen (2000)). However, there is debate as
to whether this use of heuristics re£ects cognitive de¢ciencies associated with
good moods, or more e⁄cient use of means of simplifying complex data.
   Several studies have reported that bad moods tend to stimulate people to en-
gage in detailed analytical activity, whereas good moods are associated with less
critical modes of information processing (Schwarz (1990), Petty, Gleicher, and Ba-
ker (1991), and Sinclair and Mark (1995)), so that people in good moods are more
receptive to weak as well as strong arguments (see Mackie and Worth (1991)). One
review describes the evidence as indicating that good moods cause greater reli-
ance on category information, and therefore more simplistic stereotyping (see
Bless et al. (1996)). However, Isen (2000) points out several complexities in the
interpretation of such studies owing to interacting psychological e¡ects. Bless
et al. provide evidence that good moods cause people to rely more heavily on
‘‘pre-existing knowledge structures,’’ but do not necessarily create a general de-
crease in the motivation or capacity to think e¡ectively.
   Good moods have signi¢cant positives as well. People in good moods tend to
generate more unusual associations, perform better in creative problem-solving
tasks, and show greater mental £exibility. In addition, people in good moods tend
to elaborate more on tasks involving neutral or positive (but not negative) stimu-
lus material (see, e.g., the review of Isen (2000)).
   Emotions in£uence assessments both of how favorable future prospects are
(see, e.g., Johnson and Tversky (1983) and Arkes, Herren, and Isen (1988)), and as-
sessments of risk (see, e.g., the reviews of Loewenstein et al. (2001) and Slovic et al.
(2002)). The direction of the in£uence of mood on risk assessment is complex, and
depends on the task and situation (see, e.g., the discussion of Isen (2000)).
   An important strand of the theory of a¡ective states (emotions or moods) holds
that such states provide information to individuals about the environment (see,
e.g., Frijda (1988) and Schwarz (1990)).3 A substantial body of evidence supports
an informational role of a¡ect (see, e.g., Schwarz (1990), Clore and Parrott (1991),
Wilson and Schooler (1991), and Clore et al. (1994)). Indeed, a procedure of deci-
sion making based upon feelings has been called ‘‘the a¡ect heuristic’’ by Slovic
et al. (2002).4

   3
     Consistent with this view, people often talk approvingly about making decisions based on
‘‘gut feelings,’’ good or bad ‘‘vibes,’’or doing ‘‘what your heart tells you.’’ However, ordinary con-
versation also includes criticism of bad judgments based on feelings (‘‘Marry in haste, repent
at leisure.’’).
   4
     Decisions are in£uenced not just by immediate feelings, but by the anticipation of future
feelings; see, for example, the reviews of Mellers (2000), Loewenstein et al. (2001), and Slovic
et al. (2002).
Stock Returns and theWeather                              1013

  People often attribute their feelings to the wrong source, leading to
incorrect judgments. As an example of this problem of misattribution, people feel
happier on sunny days than on cloudy days. The e¡ect of sunlight on their
judgments about happiness is reduced if they are primed by asking them
about the weather (Schwarz and Clore (1983)). Presumably this reminds
them to attribute their good mood to sunshine rather than to long-term consid-
erations.

B. Mood, Sunlight, and Behavior
  Psychologists have been documenting correlation between sunshine and beha-
vior for decades. Among other things, sunshine has been linked to tipping (Rind
(1996)), and lack of sunshine to depression (Eagles (1994)) and suicide (Tietjen
and Kripke (1994)). Most evidence suggests that people feel better when they
are exposed to more sunshine. If people are more optimistic when the sun shines,
they may be more inclined to buy stocks on sunny days. Speci¢cally, they may in-
correctly attribute their good mood to positive economic prospects rather than
good weather. This suggests that sunshine is positively correlated with stock re-
turns. Furthermore, the prediction is not that the news (as in a weather forecast)
that the day will be sunny causes an immediate and complete positive stock price
reaction. Rather, it is the occurrence of the sunshine itself that should cause
prices to move.
  An alternative to the informational perspective on a¡ect is to view feelings as
in£uencing preferences. Loewenstein (1996) reviews literature on, and models
how, ‘‘visceral factors’’ such as hunger, fatigue, sexual desire, moods, emotions,
pain, and drug cravings a¡ect preferences between di¡erent goods. Mehra and
Sah (2000) provide an analysis of the determination of stock prices when mood
a¡ects preferences.They show that small random £uctuations in preference para-
meters can cause signi¢cant volatility in prices.
  In contrast, if people are rational maximizers, there is little reason to conjec-
ture that sunshine is correlated with stock returns. It is certainly likely that
weather a¡ects economic output, particularly in industries like agriculture and
construction. However, the sunshine that occurs in one particular location is not
generally representative of the weather in an entire economy. Moreover, sun-
shine is a transitory variable. The amount of unexpected sunshine occurring to-
day is not highly correlated with the amount that will prevail one week or one
month from today. Finally, the occurrence of rain or snow should be more
strongly correlated with output than cloudiness versus sunshine. Severe weather
could hamper markets, communications, or other commercial activities. We
therefore jointly test for the incremental e¡ects on stock returns of rain, snow,
and sunshine.5

  5
    If people derive utility from both consumption and sunshine, some slight correlation be-
tween daily stock returns and sunshine may be rationally induced. However, most rational
models do not attempt to capture stocks’ sunshine risk.
1014                             The Journal of Finance

C. Mood, Sunlight, and Stock Prices
   There is already some evidence that sunshine in£uences markets. Saunders
(1993) shows that when it is cloudy in New York City, New York Stock Exchange
index returns tend to be negative. He shows that the cloudiness/returns correla-
tion is robust to various choices of stock index and regression speci¢cation.
Although this ¢nding is noteworthy, it has received little attention, possibly be-
cause of concerns about unintentional data mining. Studies that identify signi¢-
cant relationships are more likely to be published than those that ¢nd nothing, so
there will always be well-executed, published papers describing signi¢cant, but
meaningless results. Our paper helps remedy this potential concern with respect
to sunshine e¡ects by examining data from many exchanges and a more recent
time period.
   Kamstra, Kramer, and Levi (2000a) use data from several countries to show
that the Friday to Monday return is signi¢cantly lower on daylight-savings-time
weekends than on other weekends. However, such changes can a¡ect sleep pat-
terns, not just waking hours of sunlight. Furthermore, examining weather data
allows us to exploit the full sample of daily returns instead of a sample restricted
to the dates of daylight-saving-time changes.
   Kamstra, Kramer, and Levi (2000b) examine the e¡ects of seasonal shifts in
length of day in ¢ve stock markets (including markets in both the Northern and
Southern Hemispheres). They ¢nd that stock market returns are signi¢cantly re-
lated to season, and argue that this relationship arises because of the determi-
nistic shifts in the length of day through the seasons.They therefore suggest that
deterministic variations in length of day help explain the January e¡ect.6
   We examine here the relationship of stock returns to a stochastic variable, the
realization of the weather. We test the hypothesis that cloudy weather is corre-
lated with poor stock returns. We do this with a sample of 26 stock exchanges.
Using a panel rather than a long time series has three advantages.
   First, it helps identify whether the hypothesized phenomenon is pervasive.The
psychological argument for the e¡ect of sunshine should apply globally.
   Second, our recent sample allows us to examine whether this phenomenon is
still present in the years subsequent to Saunders (1993). If ¢nancial markets have
become more e⁄cient over time, it is possible that the sunshine e¡ect found by
Saunders is no longer relevant. Consistent with this notion, Saunders found that
his regression does not work well in the last subperiod of his sample (1983 to
1989). However, our study ¢nds that sunshine e¡ects are still present internation-
ally and in the United States.
   Third, the panel increases our power to detect an e¡ect. Even if sunshine af-
fects returns, we know there are many other important in£uences on any given
day. Most variation in returns will be driven by economic events and news. Given

  6
    Kramer (2000) examines another possible e¡ect of mood on stock prices. She describes
psychological evidence of predictable swings in mood over the course of the day. In a meta-
analysis of three previous empirical studies on intraday returns, Kramer argues that the evi-
dence on intraday patterns in expected stock returns is consistent with psychological pat-
terns of diurnal mood swings.
Stock Returns and theWeather                         1015

the high variability of returns, it is useful to maximize power by using a large
number of markets.

                                     II. Data
   Since we want to examine whether stock returns are correlated with cloudi-
ness, we need both stock return and weather data. We collect weather data from
the International Surface Weather Observations (ISWO) data set sold by the Na-
tional Climactic Data Center (http://lwf.ncdc.noaa.gov/oa/ncdc.html). The ISWO
data set contains detailed descriptions of the weather at 3,000 locations world-
wide from 1982 to 1997. In most ISWO locations, observations about the wind,
cloud cover, and barometric pressure are collected hourly. Since the hypothesis
that we examine relates stock returns to the amount of sunshine on a particular
day, we collect the ISWO variable that measures total sky cover (SKC). SKC
ranges from zero (clear) to eight (overcast).We calculate the average cloud cover
for each day from 6 a.m. to 4 p.m. local time for cities with stock exchanges.While
the ISWO data appear to be of very high quality, a complete record of sky cover
observations is not available for all of the cities with major stock exchanges. In
particular, SKC data for Tokyo, Hong Kong, Seoul, Lisbon, Mexico City, Toronto,
Jakarta, Frankfurt, and Wellington are not su⁄ciently complete to make data
from these cities usable.
   Of course, the daily cloud cover in any particular city is highly seasonal. For
example, winter months are associated with cloudier weather in New York City.
Many seasonal patterns have been identi¢ed in stock return data as well (e.g.,
Keim (1983)), and numerous possible causal forces exhibit annual seasonality.
To be certain that our results are driven by cloudiness rather than other seasonal
e¡ects, we therefore deseasonalize the cloud data. This provides a conservative
measure of the e¡ect of cloud cover, in the sense that we exclude any contribution
that cloud cover may make to seasonal return patterns.
   We calculate the average cloudiness value for each week of the year in each city,
and deseasonalize by subtracting each week’s mean cloudiness from each daily
mean. For example, we calculate the average value of SKCi,t for the ¢rst full
calendar week of the year for a particular city, taking an average of 80 values
(16 years times ¢ve days in the week). Then we subtract this mean from the city’s
daily SKCi,t values in the ¢rst week of each year. We denote the deseasonalized
value of SKC for city i on day t as SKCitn. The mean of SKCitn is close to zero, and
its global standard deviation is 2.19.
   To check whether our results are driven by adverse weather conditions, we in-
clude measures of raininess and snowiness in some of our regression speci¢ca-
tions.The ISWO data set contains a number of variables that describe the current
and recently passed weather at each station.We use these variables to determine
whether it is raining or has rained within the last observation period at each
station. We then average the number of observation periods for which rain is
reported per day. We perform similar calculations for snowy weather. Finally,
we deseasonalize our raininess and snowiness variables by week in the same
way that we deseasonalize the SKC variable. We denote the deseasonalized
1016                              The Journal of Finance

raininess in location i on day t RAINitn and we denote our deseasonalized snow
variable SNOWitn.
  We collect daily index returns from Datastream. All cities that have data from
at least 1988 to 1997 are included in the analysis. For most cities, we use the
market index calculated by Datastream (Datastream Global Index). However,
for several cities, other indices exist with longer time series. Thus, we collect
the following indices for the corresponding locations: Bovespa for Rio de Janeiro,
IGPA for Santiago, Hex for Helsinki, Kuala Lumpur Composite for Kuala Lum-
pur, PSE Composite for Manila, Madrid SE General for Madrid, Taiwan SE
Weighted for Taipei, and the Bangkok Composite for Bangkok.
  Some summary statistics for the sample appear in Table I. The cloudiness
measure used to calculate the means in column 3 of Table I is the original SKCit,

                                               Table I
                                   Summary Statistics
Table I displays a number of summary statistics that describe the sample. The time series for
each city listed in column 1 begins in the year listed in column 2 and ends on December 31,
1997. The variable described as Cover is total sky cover, taken from the ISWO data set. Cloud
cover ranges from zero to eight in any particular city at any particular time. Columns 3 and 4
report the mean and standard deviation of cloud cover in each city listed in column 1. Columns 5
and 6 report the mean and standard deviation of percentage returns in the local currency for
each city.

Location (1)     Begin Date (2)   Mean Cover (3)         STD Cover (4)   Mean Ret (5)   STD Ret (6)
Amsterdam             1982              5.39                 2.28           0.057          0.89
Athens                1988              3.23                 2.55           0.097          1.81
Bangkok               1982              5.51                 1.59           0.056          1.46
Brussels              1982              5.09                 2.32           0.057          0.79
Buenos Aires          1988              4.35                 2.71           0.496          4.13
Copenhagen            1982              5.35                 2.31           0.059          0.88
Dublin                1982              5.84                 1.85           0.069          1.06
Helsinki              1987              5.47                 2.39           0.044          1.12
Istanbul              1988              3.85                 2.58           0.269          2.63
Johannesburg          1982              3.00                 2.52           0.075          1.25
Kuala Lumpur          1982              6.80                 0.43           0.019          1.44
London                1982              5.74                 2.00           0.054          0.86
Madrid                1982              3.42                 2.74           0.067          1.06
Manila                1986              5.31                 1.92           0.108          1.95
Milan                 1982              4.29                 2.95           0.052          1.24
New York              1982              4.95                 2.72           0.058          0.92
Oslo                  1982              5.40                 2.37           0.068          1.39
Paris                 1982              5.25                 2.41           0.054          1.03
Rio de Janeiro        1982              5.17                 2.46           0.806          3.79
Santiago              1987              3.28                 3.07           0.112          1.00
Singapore             1982              6.70                 0.91           0.025          1.16
Stockholm             1982              5.49                 2.25           0.074          1.23
Sydney                1982              4.15                 2.40           0.048          1.09
Taipei                1982              5.55                 2.12           0.088          2.08
Vienna                1982              5.03                 2.59           0.056          0.95
Zurich                1982              5.33                 2.57           0.067          0.82
Stock Returns and theWeather                                      1017

complete with its seasonal variability. It is evident from the sample statistics that
some cities consistently experience more sunshine than other cities. Moreover,
some cities appear to have substantially higher expected returns than other
cities. This can be understood by noting that all returns are nominal returns
expressed in local currencies. Some currencies, like that of Brazil, experience
substantial in£ation, so high nominal returns are not surprising.
  A correlation matrix of both deseasonalized cloudiness, SKCitn, and of daily re-
turns appears in Table II. The cross-city correlation of returns appear below the
diagonal in the table, while the cloudiness correlations are above the diagonal.
While no measure of statistical signi¢cance is reported in the table, almost all
correlations greater in absolute value than 0.04 are signi¢cant at the ¢ve percent
level. Not surprisingly, most of the returns correlations are positive, large, and
signi¢cant. Only two estimated correlations are negative, and none are larger
than 0.65. Thus, while the global component of daily international returns is

                                            Table II
                                     Correlation Matrix
Table II displays estimated cross-city correlation coe⁄cients for the principal variables in the
analysis. Correlations of deseasonalized cloudiness appear above the diagonal, while returns
correlations appear below the diagonal. Almost all correlations with an absolute value greater
than 0.04 are statistically signi¢cant.

       Ams      Ath    Ban    Bru    Bue   Cop    Dub    Hel      Ist    Joh   Kua      Lon      Mad
Ams     &       0.06  0.01  0.74  0.01 0.29 0.12 0.12        0.08 0.02 0.01         0.43    0.07
Ath     0.12     &     0.05  0.07 0.02  0.02 0.04  0.03       0.41 0.00  0.02      0.01    0.00
Ban     0.07     0.10   &     0.02 0.01  0.00  0.02  0.02    0.04  0.01 0.02      0.03     0.01
Bru     0.46     0.16   0.12   &     0.01 0.25 0.10 0.09        0.09 0.02 0.00         0.40    0.09
Bue     0.08     0.03 0.05 0.07 &          0.01  0.01 0.00       0.02 0.02 0.02        0.01    0.01
Cop     0.29     0.15   0.11   0.28 0.05 &        0.04 0.19      0.04 0.02  0.01       0.12    0.05
Dub     0.39     0.17   0.12 0.29 0.05 0.29 &  0.02              0.00 0.00 0.00         0.37    0.03
Hel     0.32     0.07 0.10     0.26 0.05 0.33 0.25 &             0.04  0.00  0.04     0.03    0.06
Ist     0.03     0.09 0.05 0.02 0.01 0.06 0.08 0.05               &     0.01  0.00    0.03     0.01
Joh     0.19     0.11   0.08 0.14 0.00 0.17       0.19 0.27       0.06 &        0.02    0.01    0.01
Kua     0.20     0.16   0.22 0.22 0.03 0.21 0.21 0.19             0.04 0.19     &       0.00     0.02
Lon     0.63     0.10   0.08 0.35 0.09 0.24 0.40 0.28             0.05 0.21 0.21         &       0.01
Mad     0.32     0.17   0.15   0.36 0.09 0.24 0.28 0.28           0.03 0.18 0.20         0.29     &
Man     0.08     0.09 0.16     0.06 0.03 0.14     0.11 0.12       0.03 0.13 0.20         0.08     0.12
Mil     0.30     0.11   0.16   0.25 0.04 0.30 0.30 0.32           0.06 0.34 0.29         0.29     0.32
New     0.27     0.10   0.13   0.28 0.06 0.22 0.21 0.27           0.03 0.13 0.15         0.24     0.30
Osl     0.34     0.05 0.03 0.23 0.11 0.08 0.12 0.10               0.01 0.02 0.11         0.35     0.13
Par     0.46     0.13   0.11   0.36 0.07 0.28 0.34 0.30           0.02 0.22 0.24         0.39     0.30
Rio     0.55     0.14   0.08 0.47 0.09 0.26 0.33 0.27             0.04 0.18 0.19         0.47     0.37
San     0.08     0.03 0.02 0.08 0.08 0.08 0.05 0.11               0.03 0.04 0.08         0.08     0.06
Sin     0.16     0.06 0.08 0.19 0.13 0.12 0.13 0.08               0.03 0.09 0.13         0.15     0.17
Sto     0.25     0.17   0.22 0.26 0.06 0.26 0.27 0.24             0.06 0.22 0.65         0.26     0.28
Syd     0.43     0.18   0.13   0.38 0.08 0.28 0.31 0.39           0.06 0.16 0.21         0.36     0.37
Tai     0.05     0.12 0.10     0.08 0.01 0.06 0.10 0.07           0.07 0.06 0.11         0.05     0.12
Vie     0.24     0.17   0.17   0.28 0.06 0.22 0.24 0.24           0.11 0.16 0.22         0.18     0.32
Zur     0.61     0.19   0.13   0.51 0.07 0.34 0.40 0.33           0.03 0.24 0.28         0.48     0.42
1018                                       The Journal of Finance

                                                 Table IIFcontinued

        Man       Mil     New       Osl      Par      Rio     San       Sin      Sto      Syd      Tai      Vie      Zur
Ams     0.03     0.01     0.01     0.00     0.14     0.46     0.03    0.01     0.01     0.15    0.02     0.13     0.15
Ath     0.01     0.02    0.09    0.02     0.01    0.09    0.00     0.01    0.02    0.02    0.02    0.05    0.03
Ban      0.07    0.01    0.00     0.01    0.04    0.00    0.03    0.00     0.02    0.02     0.03    0.02    0.01
Bru     0.02    0.01     0.05    0.01     0.10     0.64     0.01    0.01    0.01     0.14    0.03     0.16     0.24
Bue     0.03    0.03    0.00     0.00     0.01    0.01    0.19     0.22    0.01     0.01     0.03     0.01    0.00
Cop      0.00    0.00     0.03     0.00     0.34     0.16     0.02    0.02     0.03     0.36    0.01     0.10     0.10
Dub     0.01    0.00    0.03    0.02     0.00     0.10    0.02     0.01    0.00     0.00    0.00     0.02     0.05
Hel      0.03     0.03     0.00     0.00     0.12     0.07     0.03     0.00     0.01     0.40     0.02     0.06     0.04
Ist     0.00     0.02    0.12    0.00     0.01    0.12     0.00    0.00    0.02    0.04    0.00    0.13    0.09
Joh      0.00    0.00     0.02     0.01     0.01     0.03     0.03     0.02    0.01     0.03    0.01     0.03     0.03
Kua      0.03     0.01     0.01    0.00    0.04     0.01    0.02     0.01     0.19    0.03     0.04    0.00     0.00
Lon     0.02     0.01     0.00    0.02     0.08     0.41     0.01     0.00    0.00     0.04    0.01     0.05     0.09
Mad     0.01    0.01     0.21     0.00    0.01    0.04    0.00    0.02    0.02    0.06    0.01    0.07    0.02
Man      &       0.04    0.02    0.01    0.01    0.02     0.02    0.03     0.05    0.01    0.01    0.01    0.02
Mel      0.21     &        0.00     0.00     0.01    0.01     0.02    0.03     0.03     0.00    0.01    0.00     0.02
Mil      0.06     0.19     &       0.01     0.07     0.13    0.01    0.00    0.01     0.03     0.03     0.07     0.31
New     0.01     0.05     0.12   &         0.03    0.01     0.02     0.00    0.00    0.02     0.01     0.01    0.01
Osl      0.11     0.35     0.23 0.13         &        0.11     0.02     0.01     0.00     0.40     0.02     0.01     0.06
Par      0.05     0.22     0.29 0.27         0.37     &        0.03     0.00     0.02     0.10    0.01     0.11     0.29
Rio      0.03     0.09     0.05 0.09         0.09     0.10     &       0.09    0.04     0.00    0.02    0.01     0.01
San      0.04     0.13     0.12 0.17         0.17     0.18     0.13     &       0.04    0.01     0.03    0.01    0.01
Sin      0.25     0.41     0.19   0.11       0.29     0.22     0.06     0.13     &       0.02    0.01    0.00    0.00
Sto      0.10     0.31     0.27 0.18         0.39     0.41     0.10     0.17     0.27     &       0.00     0.07     0.07
Tai      0.08     0.11     0.08  0.00       0.05     0.06     0.01     0.05     0.13     0.08     &       0.01    0.02
Vie      0.16     0.22     0.24 0.08         0.23     0.26     0.07     0.08     0.24     0.27     0.13     &        0.29
Zur      0.11     0.35     0.34 0.29         0.47     0.56     0.11     0.21     0.34     0.50     0.09     0.33     &

fairly large, there is also a large local component to each national index return.
We expect that cloudiness a¡ects the local component of an index return.
  Even after deseasonalizing cloudiness, there are many signi¢cant cross-city
correlations.The cloudiness correlations appear to be determined largely by geo-
graphy, with proximate cities exhibiting large correlations. The correlations in
returns and cloudiness present econometric problems that our test speci¢cations
will have to overcome.

                                                 III. Evidence
  This section describes the statistical results of testing the hypothesis that
cloudy weather is associated with low stock market returns. Our ¢rst set of re-
sults concerns some very simple speci¢cations estimated city by city. Simple city-
by-city speci¢cations give us an idea of the signi¢cance of cloudiness in explain-
ing returns. However, given the relatively short time series in our sample, we can-
not expect many individual city results to be statistically signi¢cant.
  Because there is a tremendous amount of variation in individual index stock
returns, to increase power we perform joint tests that employ the entire panel of
Stock Returns and theWeather                          1019

26 exchanges.We also check whether raininess or snowiness are correlated with
returns, and whether any correlation between sunshine and returns can be ex-
plained by raininess and snowiness. Finally, we examine the economic signi¢-
cance of the sunshine e¡ect with traditional measures similar to R2 and with a
simple trading strategy.

A. City-by-City Tests
  We ¢rst estimate simple regressions that are similar to those in Saunders
(1993). Speci¢cally, we estimate the parameters of the regression equation
                              rit ¼ at þ biC SKCitn þ eit                        ð1Þ
Ordinary least squares estimates of the biC coe⁄cient of this regression are re-
ported in the third column of Table III, and the associated t-statistics are re-
ported in the fourth column of Table III. The results are quite robust. Estimates
that use the original cloudiness measure (SKCit) are quite similar to those re-
ported. Regressions that replace the cloudiness measure with a variable that is
set to one when SKCit is less than one, to zero when SKCit is between one and
seven, and to minus one when SKCit is greater than seven also produce very simi-
lar results.
   It might be conjectured that it is not just cloudiness per se, but also the change
in cloudiness from the previous day that in£uences moods. While regressions of
returns on changes in cloudiness produce results that are similar to the levels
results in Table III, the levels results are slightly stronger. When both levels and
changes are included in the regression, the levels remain signi¢cant but the
changes coe⁄cient becomes insigni¢cant.
   The simple regression coe⁄cients in Table III already give an idea of the global
signi¢cance of cloudiness for returns. Four of the negative estimated coe⁄cients
are signi¢cantly di¡erent from zero using a two-tailed, ¢ve percent test. However,
Saunders (1993) argues that a one-tailed test is appropriate, since the alternative
hypothesis being examined concerns only the left tail. Using a one-tailed, ¢ve
percent test, 7 of the 26 coe⁄cients are statistically signi¢cantly negative. In con-
trast, the largest positive t-statistic is 0.83.
   We can examine the joint signi¢cance of these results with some simple non-
parametric calculations. The coe⁄cient on SKCitn , reported in column 3, is posi-
tive for 8 out of 26 cities. If the sign of each regression is an independent draw
from the binomial distribution, and if the probability of drawing a negative coef-
¢cient is 0.5, then the probability of ¢nding only 8 positive coe⁄cients out of 26
possible is 0.038. This is within the ¢ve percent level of signi¢cance for a one-
tailed test. The simple regression coe⁄cients reported in Table III suggest that
cloudiness and returns are correlated.
   However, the simple regressions may not be the most powerful way to examine
our hypothesis.The simple regressions relate the level of returns on any given day
to the percentage of cloud cover on that particular day.Thus, while a fairly cloudy
day (SKCit 5 6) may be associated with negative index returns, a very cloudy day
(SKCit 5 8) should be associated with very negative returns according to the sim-
1020                               The Journal of Finance

                                            Table III
                     Sunshine Regression and Logit Model Results
This table displays city-by-city and pooled results of estimating a regression of daily stock re-
turns on cloudiness and a logit model that relates the probability of a positive daily stock return
to cloudiness.The regression results appear in columns 3 and 4, while the logit results appear in
columns 5 to 7. An asterisk indicates statistical signi¢cance at the ¢ve percent level or greater.

                                  OLS Regression                                   Logit Model
Location             Observations (2)    biC (3)     t-Statistic (4)    giC (5)      w2 (6)   P-Value (7)
Amsterdam                  3984          0.007           1.07         0.024        2.76       0.0963
Athens                     2436           0.012            0.71         0.014        0.53       0.4649
Bangkok                    3617           0.009            0.45         0.014        0.24       0.6259
Brussels                   3997          0.018 n         3.25         0.036 n      6.75       0.0094
Buenos Aires               2565          0.030           0.98         0.019        1.60       0.2054
Copenhagen                 4042          0.002           0.30         0.002        0.02       0.8999
Dublin                     3963          0.000           0.02         0.025        2.13       0.1445
Helsinki                   2725          0.016           1.67         0.034 n      4.01       0.0452
Istanbul                   2500           0.007            0.32         0.001        0.00       0.9488
Johannesburg               3999           0.004            0.47         0.012        0.67       0.4124
Kuala Lumpur               3863           0.014            0.26         0.109        1.99       0.1586
London                     4003          0.010           1.52         0.019        1.41       0.2355
Madrid                     3760          0.011           1.60         0.015        1.41       0.2353
Manila                     2878           0.018            0.83          0.003        0.02       0.9023
Milan                      3961          0.014 n         2.03         0.021        3.69       0.0549
New York                   4013          0.007           1.28         0.035 n      8.64       0.0033
Oslo                       3877          0.018           1.92         0.025        3.31       0.0688
Paris                      3879          0.009           1.27         0.027 n      3.93       0.0474
Rio de Janeiro             2988          0.057           1.93         0.016        0.96       0.3267
Santiago                   2636           0.000            0.05         0.012        0.73       0.3935
Singapore                  3890           0.008            0.37         0.002        0.00       0.9588
Stockholm                  3653          0.014           1.54         0.025        2.89       0.0889
Sydney                     4037          0.014 n         1.96         0.020        2.16       0.1417
Taipei                     3784          0.016           0.97         0.013        0.66       0.4164
Vienna                     3907          0.013 n         2.14         0.026 n      4.11       0.0425
Zurich                     3851          0.007           1.28         0.012        0.89       0.3465
All cities (naive)        92808          0.011 n         4.49         0.020 n     43.62       0.0001
All cities (PCSE)         92808          0.010 n         3.96          F            F            F

ple regression. Saunders (1993) ¢nds that returns are negative more often on clou-
dy days than on sunny days. It is possible that while the sign of an index return on
a particular day is related to the exchange city’s cloudiness that day, the magni-
tude of the index return is not strongly related to cloudiness.
  To examine this alternative speci¢cation, we estimate logit models of the form

                                                              n
                                                      egiC SKCit
                                   Pðrit 40Þ ¼                   n                                    ð2Þ
                                                    1 þ egiC SKCit
Stock Returns and theWeather                            1021

by maximum likelihood.The ¢fth column of Table III reports the estimates of giC .
The sixth and seventh columns of Table III report each coe⁄cient’s chi-square
test of statistical signi¢cance and its associated p-value.
   While only four of the simple regression coe⁄cients are statistically signi¢-
cant in the ¢ve percent, two-tailed test, ¢ve of the logit coe⁄cients are signi¢cant
at this level. Furthermore, nine of the coe⁄cients are signi¢cant at the 10 percent
level, or equivalently, in a one-tailed test. Again, no positive estimate of giC is even
close to signi¢cant.
   In fact, 25 out of 26 estimated giC coe⁄cients are negative. Performing the same
calculation as before, if each of the signs of the giC was independently binomial
( p 5 0.5), the probability of this occuring would be 4  10  7. This is quite strong
evidence that cloudiness is correlated with returns.
   Moreover, the chi-share test statistic for the NewYork logit regression is highly
statistically signi¢cant, with an associated p-value of 0.0033.This is a notable con-
trast with the ¢nding of Saunders (1993) that the sunshine e¡ect is insigni¢cant in
the last subperiod of his sample (1983 to 1989). Saunders concluded that the sun-
shine e¡ect may be of purely historical interest. Splitting the New York logit re-
gression into similar subperiods, the logit coe⁄cient estimate calculated with our
data for the eight-year period from 1982 to 1989 is  0.0136 with an associated chi-
square statistic of 0.68 (p-value 5 0.4081). However, the logit coe⁄cient estimate
for the eight-year period from 1990 to 1997 is  0.0578 with an associated chi-
square statistic of 11.38 (p-value 5 0.0007). Thus, although we replicate Saunders’
¢nding that the weather e¡ect was weak in the 1980s, the sunshine e¡ect appears
most strongly in New York during the 1990s. Because of sampling noise, eight
years is a short period of time to measure these e¡ects. Thus, the results are not
inconsistent with a stable sunshine e¡ect through the entire period.

B. Joint Tests
  While the city-by-city results reported above strongly suggest that stock re-
turns are correlated with cloudiness, we can use the entire data set to make more
de¢nitive statements about the statistical signi¢cance of the cloudiness e¡ect.
We report the results of several joint (across cities/indices) tests of signi¢cance
in this section.
  The ¢rst joint tests can be considered simple generalizations of the regressions
described above. We estimate one regression with the simple speci¢cation of
equation (1) with all of the data from each city.We refer to this as a pooled regres-
sion. In particular, we estimate a pooled regression of the form
                               rit ¼ a þ bC SKCitn þ eit ;                          ð3Þ
where now the parameters a and bC are constrained to be the same across mar-
kets. In this simple pooled speci¢cation, we assume that the error terms, eit, are
i.i.d. This speci¢cation does not adjust for any contemporaneous correlations
across the error terms of di¡erent indices, nor does it adjust for any autocorrela-
tion among a particular index’s errors. Contemporaneous correlation across in-
dex-speci¢c error terms almost certainly exists, given the correlations in Table II.
1022                          The Journal of Finance

  The results of the simple pooled regression appear in the penultimate line of
Table III. The bC coe⁄cient estimate is  0.011, which is approximately the mean
of the city-by-city estimates. The associated t-statistic is  4.49, which is highly
statistically signi¢cant.
  We also perform a pooled test with the logit model described by equation (2).
Again, we simply concatenate the data from each city, resulting in one sample
of 92,808 observations, 53.2 percent of which are positive. The estimate of gc is
 0.02, approximately the mean of the city-by-city results. Moreover, the chi-
square test of statistical signi¢cance is 43.62, which is very statistically signi¢-
cant. However, these test statistics are also based on the dubious assumption that
each observation is i.i.d.
  To allow for violations of the assumption that each error term is i.i.d., we esti-
mate a city-speci¢c ¢xed e¡ects model of the form
                              rit ¼ ai þ bC SKCitn þ eit ;                       ð4Þ
with panel corrected standard errors (PCSE). Our PCSE speci¢cation allows eit
to be contemporaneously correlated and heteroskedastic across cities, and auto-
correlated within each city’s time series. We estimate bC to be  0.010 with an
associated asymptotic z-statistic of  3.96. These estimates are quite close to
the naive pooled estimates discussed above, so adjusting for the correlations in
the panel data does not reduce the power of the inference very much. Adjusting
the logit model for panel correlations is signi¢cantly more complicated than ad-
justing the simple regression. Therefore, we estimate a ¢xed e¡ects linear prob-
ability model with PCSE of the form
                         Iðrit 40Þ ¼ aP;i þ bP SKCitn þ eit ;                    ð5Þ
where I(rit40) is an indicator variable that is one when returns are positive. Es-
timating this regression yields an estimate of bP of  0.005 with an associated t-
statistic of  6.07. Again, adjusting for contemporaneous correlation and hetero-
skedasticity across panels and autocorrelation within panels has little e¡ect on
the inference.
   The sunshine e¡ect has been criticized by Trombley (1997) on the grounds that
the results documented for New York weather and returns in Saunders (1993) are
not statistically signi¢cant in each month of the year and subperiod of the data
(a criterion which we regard as inappropriate). Trombley also ¢nds that average
returns do not appear to be a monotonic function of cloudiness in Saunders’data.
In unreported tests that employ our entire panel of data, we ¢nd that average
returns are almost monotonically decreasing in cloudiness and that the e¡ect of
cloudiness on returns is negative in all 12 months of the year and signi¢cantly
negative in one-sided tests in 6 of the 12 months. Overall, our ¢nding that sunshine
is statistically signi¢cantly correlated with daily returns appears quite robust.

C. Controlling forAdverseWeather
  As discussed above, it is possible that sunshine is just a proxy variable for other
weather conditions, such as lack of rain or snow, that may be correlated with
Stock Returns and theWeather                              1023

stock returns.We examine this hypothesis by measuring raininess and snowiness
and including adverse weather conditions in the regressions. The regressions
that we estimate and report in Table IV take the form
                     rit ¼ ai þ biC SKCitn þ biR RAINitn þ biS SNOWitn þ eit ;               ð6Þ
where the variables are measured as de¢ned in Section II. Table V reports esti-
mates of logit models analogous to those described in the previous subsection.
  The results in Tables IVand V indicate that the sunshine e¡ect is not explained
by other weather conditions. In Table IV, only 9 of the 26 sky cover coe⁄cients
reported in column 2 are positive. In the regressions without raininess and snow-
iness (in Table III), 8 of the coe⁄cients are positive. Similar to the regressions

                                             Table IV
      Sunshine Regressions Controlling for Other Weather Conditions
This table displays city-by-city and pooled results of estimating a regression of daily stock re-
turns on cloudiness, raininess, and snowiness. Column 2 contains the coe⁄cient estimate for
cloudiness, and column 3 contains the associated t-statistic. Columns 4 and 5 contain coe⁄cient
estimates for raininess and snowiness. An asterisk indicates statistical signi¢cance at the ¢ve
percent level or greater.

  Location (1)                   biC             t-Statistic            biR                biS
Amsterdam                       0.005              0.69             0.064              0.075
Athens                           0.014               0.77             0.295              0.393
Bangkok                          0.014               0.63             0.566               0.392
Brussels                        0.011              1.76             0.219 n             0.123
Buenos Aires                    0.022              0.64             0.580               2.226
Copenhagen                      0.001              0.10             0.193               0.264
Dublin                           0.001               0.08             0.018               0.111
Helsinki                        0.016              1.51              0.319              0.397
Istanbul                         0.008               0.33             0.340               2.575
Johannesburg                     0.003               0.28              0.090               0.044
Kuala Lumpur                     0.021               0.38             0.206              0.023
London                          0.009              1.14             0.052              0.019
Madrid                          0.011              1.42              0.031              0.411
Manila                           0.015               0.63              0.234              2.318
Milan                           0.015 n            1.99              0.342              1.026
New York                        0.002              0.31             1.929 n            1.555
Oslo                            0.019              1.76              0.035               0.078
Paris                           0.012              1.53              0.279               0.378
Rio de Janeiro                  0.067 n            2.16              1.135               2.805
Santiago                         0.001               0.17             0.409              1.154
Singapore                        0.007               0.32              0.095              0.377
Stockholm                       0.012              1.22              0.094              0.594
Sydney                          0.013              1.51             0.073               0.459
Taipei                          0.020              1.13              0.212               3.378
Vienna                          0.013 n            2.00              0.005               0.009
Zurich                          0.002              0.31             0.101              0.015
All cities (naive)              0.010 n            3.94             0.058               0.076
All cities (PCSE)               0.009 n            3.47             0.065               0.067
1024                              The Journal of Finance

                                            TableV
       Sunshine Logit Models Controlling for Other Weather Conditions
This table displays city-by-city and pooled results of estimating a logit model that relates the
probability of a positive daily stock return to cloudiness, raininess, and snowiness. Column 2
contains the coe⁄cient estimate for cloudiness, and column 3 contains the associated w2 -statis-
tic. Columns 4 and 5 contain coe⁄cient estimates for raininess and snowiness. An asterisk in-
dicates statistical signi¢cance at the ¢ve percent level or greater.

  Location (1)                giC (2)            w2 (3)             giR (4)              giS (5)
Amsterdam                     0.028              3.04               0.133                0.023
Athens                        0.014              0.51               0.357               2.819
Bangkok                       0.007              0.05              0.810               0.207
Brussels                      0.019              1.42              0.565 n              0.474
Buenos Aires                  0.012              0.51              0.254               0.272
Copenhagen                     0.002              0.01              0.435                0.103
Dublin                        0.012              0.40              0.225                0.739
Helsinki                      0.038 n            4.07               0.606               0.130
Istanbul                       0.003              0.02              0.557                1.441
Johannesburg                  0.016              1.05               0.166                0.298
Kuala Lumpur                  0.103              1.74              0.277                0.310
London                        0.007              0.16              0.330               0.034
Madrid                        0.018              1.59               0.300                0.989
Manila                         0.005              0.04              0.151                0.605
Milan                         0.021              2.94               0.180               1.742
New York                      0.025 n            4.14              4.943 n              2.636
Oslo                          0.024              2.45               0.021               0.120
Paris                         0.031 n            4.46               0.242                2.060
Rio de Janeiro                0.022              1.83               0.945                0.472
Santiago                      0.009              0.34              1.074               1.615
Singapore                     0.008              0.05               0.417               0.467
Stockholm                     0.015              0.85              0.988               1.010
Sydney                        0.019              1.48              0.053                0.386
Taipei                        0.015              0.77               0.103                2.542
Vienna                        0.030 n            4.51               0.229                0.776
Zurich                         0.001              0.00              0.289 n              0.112
All cities (naive)            0.017 n           28.86              0.190 n              0.128

without raininess and snowiness, 3 of the coe⁄cients are signi¢cantly negative at
the ¢ve percent level, but none of the coe⁄cients is close to signi¢cantly positive.
By comparison, raininess and snowiness have 12 and 13 positive estimated coe⁄-
cients in columns 4 and 5, respectively. Two of the raininess and 1 of the snowi-
ness coe⁄cients are statistically signi¢cant.
   While the city-by-city results in Table IV suggest that the sunshine e¡ect is in-
dependent of raininess and snowiness, we can design more powerful tests of the
adverse weather explanation by considering all cities’ returns jointly.The last two
lines of Table IV report the results of an all city pooled regression and a ¢xed-
e¡ects PCSE regression analogous to those regressions described in the previous
subsection. In both regressions, the coe⁄cient on sky cover is signi¢cantly nega-
tive, with t-statistics of  3.94 and  3.47. However, raininess and snowiness are
Stock Returns and theWeather                          1025

economically and statistically insigni¢cant in both speci¢cations. Overall, the
results of Table IV do not support the other-weather-condition explanation of
our results.
  The binary results in Table V are not favorable to the adverse weather hypoth-
esis either. As in the previous section, the logit model estimates reported in Table
V relate the sign of the return in city i on day t to the weather in that city on that
day. Speci¢cally, the models we estimate are of the form
                                              n         n           n
                                   egiC SKCit þgiR RAINit þgiS SNOWit
                     Pðrit 40Þ ¼               n           n          n :          ð7Þ
                                 1 þ egiC SKCit þgiR RAINit þgiS SNOWit

Looking at the city-by-city results, only 4 of the sky cover coe⁄cient estimates
reported in column 2 are positive, while 4 of the estimates are signi¢cantly nega-
tive. While 3 of the raininess coe⁄cients reported in column 4 are signi¢cantly
negative, 12 of the raininess coe⁄cients are positive. None of the snowiness coef-
¢cients in column 5 is signi¢cant, and 16 of them are positive. The city-by-city
logit results con¢rm that sunny days tend to coincide with positive returns.
   To assess the joint signi¢cance of the city-by-city logit estimates, we again es-
timate a naive pooled logit model and a ¢xed-e¡ects PCSE linear probability
model. The results of the pooled logit model appear in the ¢nal row of Table V.
The pooled estimate of the logit coe⁄cient on sky cover is  0.017, which is close
to the mean of the city-by-city estimates. The associated chi-square statistics is
28.86, which indicates that the coe⁄cient estimate is very statistically signi¢-
cant.The coe⁄cient estimate for raininess is  0.190, and it is statistically signif-
icant. The estimate for snowiness is positive and insigni¢cant.
   As in the case without adverse weather controls, the ¢xed-e¡ects PCSE linear
probability model is consistent with the pooled logit model. The coe⁄cient esti-
               n
mate for SKCi;t  is  0.0043, with an asymptotic z-statistic of  5.01.The coe⁄cient
                      n
estimate for RAINi;t    is  0.049, with a z-statistic of  3.06, and the coe⁄cient es-
                    n
timate for SNOWi;t is positive and insigni¢cant. Again, even after controlling for
other weather conditions, sunshine is strongly signi¢cantly correlated with both
the sign and the magnitude of returns.

D. Ability of Sunlight to Predict Returns
  With the parameter estimates reported in Table III, we now consider the pro¢t-
ability of trading strategies based upon the sunshine e¡ect, and whether morn-
ing sunlight can predict returns for the day. We use the coe⁄cient of the simple
pooled regression,  0.011, as our estimate of the sunshine e¡ect.The mean daily
return across all countries is 0.103, with a standard deviation of 1.58. We know
that the cloudiness variable ranges from zero to eight, so the di¡erence in ex-
pected return between a completely overcast day and a sunny day is 0.088 or about
nine basis points.While nine basis points is approximately how much the markets
go up on an average day, it is only about ¢ve percent of the standard deviation of
daily returns. Consistent with this calculation, the R2 of the naive pooled regres-
sion reported in Table III is 0.02 percent, a very low number.
1026                          The Journal of Finance

   It is, of course, not reasonable to expect the explanatory power of sunshine to
be large. Many shocks a¡ect daily stock returns, such as real news about global,
national, and local fundamentals. Unless the market is grotesquely ine⁄cient,
fundamental news must have a large e¡ect on returns.
   Rather than focusing on R2, we consider whether a portfolio strategy based on
weather trading can signi¢cantly increase the Sharpe ratio of a hypothetical in-
vestor. We employ a Britten-Jones (1999) test of the mean-variance e⁄ciency of a
simple global market portfolio, and the global portfolio combined with a weather-
based strategy. The Britten-Jones test regresses a vector of ones on portfolio re-
turns. When more than one set of portfolio returns are used as dependent vari-
ables, a mean-variance e⁄cient portfolio will be signi¢cantly related to the
vector of ones and all other portfolios will be unrelated to the vector of ones.
The intuition behind the test is that a vector of ones represents the ideal asset
returnFthe returns are positive with no variability. Put di¡erently, the vector
of ones represents the returns of an asset with an in¢nite Sharpe ratio, having a
mean of one and a variance of zero.The regression ¢nds the combination of poten-
tially mean-variance e⁄cient portfolios that most closely approximates this ideal
asset return.
   Our global market portfolio is the equal-weighted portfolio of all cities’ returns
in local currencies. To construct our trading strategy returns, we average SKCi,t
for each city from 5 a.m. to 8 a.m. each morning.We then deseasonalize the morn-
ing SKC variable in the same way that we deseasonalize the previous measure of
SKC (as described in Section II). Finally, we calculate the equal-weighted average
return of cities with positive deseasonalized morning SKC and the equal-
weighted average return of cities with negative deseasonalized morning SKC.
We consider strategies that are long indices with sunny cities, short indices with
cloudy cities, and both long the sunny indices and short the cloudy indices. The
results of our test appear in Table VI.
   Table VI shows that in the absence of transaction costs investors can improve
their Sharpe ratios by trading on the weather. In the ¢rst model, the returns of
the sunny strategy are given almost the same weight in the mean-variance e⁄-
cient portfolio as the returns of the global market portfolio. The weight on the
sunny strategy returns is statistically signi¢cant, with a t-statistic of 3.39. In
the second model, the mean-variance e⁄cient portfolio is clearly short the cloudy
portfolio. The t-statistics of the weight of the cloudy portfolio is  3.22. The re-
sults of the third model are somewhat ambiguous, presumably because of a high
degree of multicollinearity in the regression. However, the fourth model again
implies that trading on the weather can increase a Sharpe ratio. The strategy of
buying the sunny portfolio and selling the cloudy portfolio is clearly a substantial
part of the global mean-variance e⁄cient portfolio.

E. Accounting for Transaction Costs
   To determine whether exploiting the sunshine e¡ect can increase a Sharpe ra-
tio after accounting for reasonable transaction costs, we propose a strategy that
requires fewer trades than the strategy described above, and we calculate the
Stock Returns and theWeather                                   1027

                                            TableVI
                     Tests of Weather-based Trading Strategies
This table displays results of tests of the mean-variance e⁄ciency of a global equity portfolio.
The global equity portfolio return is the equal-weighted average of local currency returns of the
26 stock exchanges in the sample. Tests are performed within the framework of Britten-Jones
(1999).We estimate regressions of the form
                                     1 ¼ bm rmt þ bs rst þ ut
where rmt is the return on the global equity portfolio and rst is the return to a weather-based
strategy on day t. The weather-based strategies that we consider are: the sunny strategy, in
which we invest an equal amount in each exchange for which morning total sky cover SKCit is
negative on day t; the cloudy strategy, in which we invest an equal amount in each exchange for
which morning SKCit is positive on day t; and the sunny^cloudy strategy, in which we go long
the sunny strategy and short the cloudy strategy. Morning SKCit is measured from 5 a.m. to 8
a.m. local time in each city on each day. A statistically signi¢cant coe⁄cient for a return series
in one of our models implies that the return series is an important component of a mean-var-
iance portfolio. The t-statistics are in parentheses.

Model (1)          Market (2)           Sunny (3)           Cloudy (4)         Sunny^Cloudy (5)
Model 1               13.53               13.52
                      (2.92)              (3.39)
Model 2               43.11                                      16.16
                      (7.59)                                    (  3.22)
Model 3               23.88                9.35                    6.19
                      (1.45)              (1.24)                (  0.66)
Model 4               27.03                                                            7.97
                     (10.72)                                                          (3.45)

returns to our strategy net of costs. In particular, we calculate the returns to a
strategy that is long the market index of each city for which the average cloud
cover variable between 5 a.m. and 8 a.m. is between zero and four, and short the
city’s index otherwise. Each city for which the strategy stipulates a long position
receives a weight of 1/n (where n is the number of cities considered by the strat-
egy) while each city for which the strategy stipulates a short position receives a
weight of  1/n. Since the number of cloudy cities varies from day to day, this trad-
ing rule requires a positive net investment on some days and a negative net in-
vestment on other days. If the trading rule implies that the position in a
particular city changes from a long to a short position or from a short to a long
position, we subtract transaction costs from the return of the city on that day.
After calculating the returns to this strategy for various levels of costs, we ask
whether the strategy return should receive positive weight in a mean-variance
e⁄cient portfolio using the Britten-Jones test described above.
  Performing the test with costs of one basis point per transaction (two points
for changing a position from long to short or from short to long), the Britten-
Jones test assigns a weight of 34.02 (t 5 11.20) to the market portfolio and a weight
of 20.21 (t 5 4.29) to the weather strategy. For comparison, with no transaction
costs the Britten-Jones test gives optimal weights of 36.11 (t 5 11.92) and 26.07
You can also read